Upload
others
View
7
Download
0
Embed Size (px)
Citation preview
Class Projects • Each team must make a 15minute presentation on
December 18 (1:30-4:30pm). • All team members must be present and participate. • Early oral reports can be arranged and are encouraged!
Please email me if you want to do this. • The written report must be emailed to
[email protected] by 6am on December 19. Late reports are not allowed.
12/1/2020 Intro to Path Integral MC 1
12/1/2020 Intro to Path Integral MC 2
Motivation for Path Integral MC
• We need to go beyond classical simulations since the microscopic world is quantum mechanical. – PIMC makes nice generalization of classical
simulations related to polymers,…. – and leads to understanding of bose condensation,
superfluidity, exchange … • Methods of classical simulation carry over to quantum
simulation. – BUT only for thermodynamic properties of bosonic
systems • Details given in Rev. Mod. Phys. 67, 279 (1995),
12/1/2020 Intro to Path Integral MC 3
Imaginary Time Path Integrals
12/1/2020 Intro to Path Integral MC 4
PIMC Simulations • We do Classical Monte Carlo simulations to evaluate
averages such as:
• Quantum mechanically for T>0, we need both to generate the distribution and do the average:
• Simulation is possible since the density matrix is positive.
<V >= 1
ZdRV (R)e−βV ( R)∫ with β = 1/ (kBT )
<V >= 1Z
dR V (R)ρ R;β( )∫ρ R;β( ) = diagonal density matrix
12/1/2020 Intro to Path Integral MC 5
Notation • Individual coordinate of a particle ri
• All 3N coordinates R= (r1,r2, …. rN)
• Total potential energy V(R)
• Kinetic energy
• Hamiltonian
−λ ∇i
2
i=1
N
∑ where λ ≡ !2
2m
ˆ ˆ ˆH T V= +
12/1/2020 Intro to Path Integral MC 6
The thermal density matrix • Find exact many-body
eigenstates of H. • Probability of
occupying state α is exp(-βEα)
• All equilibrium properties can be calculated in terms of thermal o-d density matrix
• Convolution theorem relates high temperature to lower temperature.
2
ˆ-
*
1 2 1 2
1 1 2
off-diagonal density matrix:
operator notation
ˆ
( ; ) ( )
ˆ
( , '; ) ( ') ( )
( , '; ) 0 (without statistics)( , ; )
= ' ( , '; ) ( ', ;
E
H
E
H E
R R e
e
R R R R e
R RR R
dR R R R R
α
α
α α α
βα
α
ββ
βα α
α
φ φ
ρ β φ
ρ
ρ β φ φ
ρ βρ β β
ρ β ρ β
−
−
=
=
=
=
≥+ =
∑
∑
1 2 1 2
2
ˆ ˆ ˆ-( ) - -or with operato
)
rs : H H He e eβ β β β+ =
∫
12/1/2020 Intro to Path Integral MC 7
Trotter’s formula (1959) • We can use the effects of operators separately as long as we take small enough time steps. • M is number of time slices. • τ is the “time-step”
• We now have to evaluate the density matrix for potential and kinetic matrices by themselves:
• Do by FT’s
• V is “diagonal”
• Error at finite n is roughly: comes from communtator
ρ̂ = e−β (T!+V! )
ρ̂ = lim M→∞ e−τ T!e−τV!⎡⎣
⎤⎦
M
τ = β / M
r e−τ T! r ' = 4πλτ( )−3/2e− r−r '( )2 /4λτ
r e−τ V! r ' = δ (r − r ')e−τV (r )
2ˆ ˆ,
2T V
eτ ⎡ ⎤− ⎣ ⎦
12/1/2020 Intro to Path Integral MC 8
Evaluation of kinetic density matrix
r e−τ T! r ' = φα* (r)φα (r ')e−τTα
α∑
In PBC eigenfunctions of T̂ = 1
Ωe− i"k"r
and eigenvalues are λk 2
r e−τ T! r ' = 1
Ωe− i"k"r ei
"k"r 'e−τλk2
k∑
convert to an integral
r e−τ T! r ' = 1
2π( )3 dkei"k ( "r '−"r )−τλk2
∫ = 4πλτ( )−3/2e− r−r '( )2 /4λτ
Danger: makes assumption about boundaries and statistics.
12/1/2020 Intro to Path Integral MC 10
• We sample the distribution:
Where the “primitive” link action is:
• Similar to a classical integrand where each particle turns into a “polymer.” – K.E. is spring term holding polymer together. – P.E. is inter-polymer potential.
• Trace implies R1=Rm+1 ð closed or ring polymers
e− S(Ri ,Ri+1 ;τ )
i=1
M
∑/ Z with Z = dR
1∫ ...dRMe− S(Ri ,Ri+1 ;τ )
i=1
M
∑
S(R0,R1;t) = − 3N
2ln 4πλτ( ) + (R0 − R1)2
4λτ+ τ2 V(R0 )+ V(R1)⎡⎣ ⎤⎦
Using this for the density matrix.
12/1/2020 Intro to Path Integral MC 11
“Distinguishable” particles • Each atom is a ring
polymer; an exact representation of a quantum wavepacket in imaginary time.
• Trace picture of 2D helium. The dots represent the “start” of the path (all points are equivalent)
• The lower the real temperature, the longer the “string” and the more spread out the wavepacket.
12/1/2020 Intro to Path Integral MC 12
Different schemes to picture PIs.
Space filling picture Fourier smoothed trace
World line picture Discretized trace
12/1/2020 Intro to Path Integral MC 13
Main Numerical Issues of PIMC • How to choose the action. We don’t have to use the
primitive form. Higher-order forms cut down on the number of slices by a factor of 10. We can solve the 2-body problem exactly.
• How to sample the paths and the permutations. Single slice moves are too slow. We move several slices at once. Permutation moves are made by exchanging 2 or more endpoints.
• How to calculate properties. There are often several ways of calculating properties such as the energy.
If you use the simplest algorithm, your code will run 100s or 1000s of times slower than necessary.
Calculations of 3000 He atoms can be done on a workstation-- if you are patient.
Details see: RMP 67, 279 1995.
12/1/2020 Intro to Path Integral MC 14
Calculating properties • Procedure is simple: write down observable:
• Expand density matrix into a “path”:
• Density, density-density, …. the potential energy are diagonal operators. Just take average values as you would classically.
• All time slices are the same – can use all for averages.
ˆˆ' ' 'HdRdR R O R R e RO
Z
β−
< >=∫
path average
path average
ˆ '
ˆ ( ) for "diagonal operatorsk
O R O R
O O R
< >=
< >=
12/1/2020 Intro to Path Integral MC 15
Calculation of Energy • Thermodynamic estimator: differentiate partition function
Problem: variance diverges as small time step. • Virial Estimator: differentiate in “internal coordinates”
does not diverge at small time steps (Herman, Berne)
E = − dZZdβ
= 1Z
dR∫ e−S dSdβ
⎡
⎣⎢
⎤
⎦⎥ =
dSk
dτpath
dSdτ
= dUdτ
+ 3N2τ
−R − R '( )2
4λτ 2
Potential NI-KE deviation from centroid .force
Evirial =
dUdτ
+ 3N2β
+ 12τ
(Ri −C) ⋅∇iU
Potential n*NI-KE spring energy
12/1/2020 Intro to Path Integral MC 19
Examples of distinguishable particle calculations
• Solid H2: work of Marcus Wagner, DMC
• Wigner crystal: 3D Matt Jones, DMC 2D Ladir Candido, P. Phillips, DMC
12/1/2020 Intro to Path Integral MC 20
Example: Solid H2
Solid molecular hydrogen is a very quantum solid
KE=69K Tt = 13.8K • Below Tt interface between
solid and gas. • Top layer is at a lower density,
more delocalized and interesting quantum effects
• Normally freezing at surface is depressed by 10%.
• In H2 it is depressed by 100%.
r 2 1/2= 0.21rNN
12/1/2020 Intro to Path Integral MC 21
Simulation is of 5 layers Each layer is 30 H2 Hard wall on left Top layer melts around
7K. Very fluffy top layer. New layer above 6K Wagner, DMC, JPTP 102,275
(1996). Z (A) T=6K
Layer Structure of Solid H2
12/1/2020 Intro to Path Integral MC 22
Snapshots of H2 density
Layer 0 Layer 1 Layer 2
T=6K T=7.5K T=10K
Gas
Liquid
Vacancy
crystal
12/1/2020 Intro to Path Integral MC 24
Melting of the 3D Wigner Crystal • PIMC with Boltzmann
statistics • Phase boundary
determined with free energy calculation
• Sudden change from pressure melting to thermal melting.
• Lindemann law is inaccurate
• Melting is first order with no volume change
Jones & DMC PRL 1996
superfluid
12/1/2020 Intro to Path Integral MC 29
• There exists an “exact link action” :
• The “primitive” link action is:
• We often define the exact “inter-action” as:
potential term = total - kinetic term (topological)
( )i i 1 i i 1
1 1
S(R ,R ; ) S(R ,R ; )
1 M
ˆ- Hi i 1 i i 1S(R
/ whe
,R ;
re dR ...d
)
R
ln R e RM M
i ie Z Z e
τ
τ τ
τ
+ += =
+
− −
+ = −
∑ ∑= ∫
S(R0,R1;t) = − 3N
2ln 4πλτ( ) + (R0 − R1)2
4t+ τ
2V(R0 )+ V(R1)⎡⎣ ⎤⎦
Improved Actions
U (Ri , Ri+1;τ ) = S(Ri , Ri+1;τ ) − S0 (Ri , Ri+1;τ )
12/1/2020 Intro to Path Integral MC 30
Improved Action
• If we make better actions, we can drastically cut down on the number of time slices. • This saves lots of time, because the number of variables to integrate over is reduced • but also because the correlation time of the walk is reduced since “polymers” are less entangled • Possible approaches to better actions:
– Harmonic approximation – Semi-classical approximation (WKB) – Cumulant approximation – Pair-product approximation
• Improved actions are also used in lattice gauge theory: the “perfect action.”
12/1/2020 Intro to Path Integral MC 32
Examples for 2 particles • Exact action • Cumulant action • Primitive action • WKB
Hydrogen atom
• Exact action
• Cumulant action
• Primitive action
• WKB
Exact action is smoother than the primitive form
WKB does not converge
He-He action
12/1/2020 Intro to Path Integral MC 36
Harmonic Approximation • We can exactly calculate the action for a harmonic oscillator. It
is just a shifted Gaussian. • In the neighborhood of (R,R’) let’s approximate the potential
by a harmonic one. • Reasonable if the potential is really harmonic within a thermal
wavelength. (for example in the high temperature limit)
• R* is an arbitrary place to evaluate the potential. If we choose it to be one of the end-points we get the Wigner-Kirkwood approximation.
• Bad idea for realistic potentials because expansion does not converge uniformly. Problem is at small r. Look at derivatives.
U H (RO , RF ;τ ) = τV (R*) + τ 2λ6 ∇2V (R*)- τ 3λ
12 ∇V (R*)⎡⎣ ⎤⎦2
- τ12 RF − RO( )∇∇V (R*) RF − RO( )
for LJ r−5 = r−12 + r−14 + r−26 + r−14 + ....
12/1/2020 Intro to Path Integral MC 37
Cluster action • For spherically symmetric pair potentials. • Find the action for a reduced subset of particles exactly and
put together to get a many-body action.
• Generalization of T=0 of the Jastrow wavefunction to finite temperatures.
• At finite T, it is the off-diagonal terms that are important.
e−U ( R0 ,RF ;τ ) = exp − dt vi< j∑ (rij (t))
0
τ
∫⎡
⎣⎢⎢
⎤
⎦⎥⎥
RW
= exp − dtv(rij (t))0
τ
∫⎡
⎣⎢⎢
⎤
⎦⎥⎥i< j
∏RW
≈ exp − dtv(rij (t))0
τ
∫⎡
⎣⎢⎢
⎤
⎦⎥⎥i< j
∏RW
• take the uncorrelated average: • This is now a 2 particle problem.
12/1/2020 Intro to Path Integral MC 46
• We need to perform integrals over the distribution:
• Where the exact link action is kinetic and potential energy:
• Similar to a classical collection of ring “polymers”. • 3NM degrees of freedom. 64 He atoms*40 slices=2560
classical particles • Available classical methods are Monte Carlo or Molecular
Dynamics. (in fact many different MC methods)
e− S(Ri ,Ri+1 ;τ )
i=1
M
∑/ Z
S(R0,R1;t) = − 3N
2ln 4πλτ( ) + (R0 − R1)2
4τ+ U(R0,R1)
Path Integral Sampling Methods
12/1/2020 Intro to Path Integral MC 47
How to sample a single slice. • pdf of the midpoint of the
bridge:(a pdf because it is positive, and integrates to 1)
• For free particles this is easy-a Gaussian distribution
PROVE: product of 2 Gaussians is a Gaussian.
• Interaction reduces P(R) in regions where spectator atoms are.
• Better is correlated sampling: we add a bias given by derivatives of the potential (for justification see RMP pg 326)
• Sampling potential Us is a smoothed version of the pair action.
P(Rt /2 ) =R0 e− tH /2 Rt /2 Rt /2 e− tH /2 Rt
R0 e− tH Rt
Rt /2 =1
2R0 + Rt( ) +η
σ 2 = λt / 2 = η2
Rt /2 =1
2R0 + Rt( ) + λt∇Us(Rt /2
0 )+η
σ2
= λt / 2I + (λt)2!∇"∇Us(Rt /2
0 ) =!η"η
Us(R) = sampling potential
12/1/2020 Intro to Path Integral MC 48
Lévy construction • How to generate a random
walk by starting in the middle. • So you don’t fall into Zeno’s
paradox. • Construct a whole path by
recursively sampling bridges – Midpoint – Midpoint of midpoints – Etc. – Stop when you are at the
desired level of precision.
P(Rt /2 ) =R0 e− tH /2 Rt /2 Rt /2 e− tH /2 Rt
R0 e− tH Rt
Rt /2 =1
2R0 + Rt( ) +η
σ t2 = λt / 2 = η2
12/1/2020 Intro to Path Integral MC 49
PIMC Sampling considerations • Metropolis Monte Carlo that moves a single
variable is too slow and will not generate permutations.
• We need to move many time slices together • Key concept of sampling is how to sample a
“bridge”: construct a path starting at R0 and ending at Rt.
• How do we sample Rt/2? GUIDING RULE. Probability is:
• Do an entire path by recursion from this formula.
• Related method: fourier path sampling.
P(Rt /2 ) =R0 e− tH /2 Rt /2 Rt /2 e− tH /2 Rt
R0 e− tH Rt
12/1/2020 Intro to Path Integral MC 50
Multi-level sampling We need to sample several links at once. Why? • Polymers move slowly as number of links increase. • Maximum moving distance is order: • Calculate how much CPU time it takes the centroid of a
single particle’s path to move a given distance • Scales as M3 . Hence doubling the number of time slices
will slow down code by a factor of 8! Eventually you get into trouble.
• (also shows why good actions help) • Permutations/windings will not get accepted easily
because pair permutations need to have the path move as well.
λτ
12/1/2020 Intro to Path Integral MC 53
Bisection method 1. Select time slices
0
ß
3. Sample midpoints
4. Bisect again, until lowest level
5. Accept or reject entire move
2. Select permutation from possible pairs, triplets, from:
ρ(R, PR ';4τ )
R’
R
12/1/2020 Intro to Path Integral MC 54
Multilevel Metropolis/ Bisection Sample some variables
Continue? Sample more variables Continue? Finally accept entire move.
• Introduce an approximate level action and sampling.
• Satisfy detailed balance at each level with rejections (PROVE)
• Only accept if move is accepted at all levels.
• Allows one not to waste time on moves that fail from the start (first bisection).
Ak s→ s '( ) = min 1,
Tk (s '→ s)π k (s ')π k−1(s)Tk (s→ s ')π k (s)π k−1(s ')
⎡
⎣⎢
⎤
⎦⎥
12/1/2020 Intro to Path Integral MC 56
MC versus MD sampling • MD can be used BUT basic algorithm is not ergodic
because spring terms do not exchange “pseudoenergy” with the other degrees of freedom.
(in Fermi-Ulam-Pasta experiment, slightly anharmonic chains never come into equilibrium.)
• Coupled themostats are introduced to solve this problem--but requires some detailed tinkering to make it work in many cases.
• Basic problem with MD-cannot do discrete moves needed for bose/fermi statistics
• An advantage of MD is that multiparticle moves are natural-allows fast computation of energy and forces within LDA.
• Little systematic comparison for a realistic systems. • Development of “worm algorithms” for lattice systems.